Study Of Control Method For Quadruped Long Jump Based On Deep Reinforcement Learning

Quadrupeds have a very strong adaptive capacity in unstructured environments.Quadruped robots draw on the body structure of quadrupeds and inherit the strong adaptive ability of quadrupeds.This makes the quadruped robot’s adaptability to complex environments a huge advantage over wheeled robots,with very broad application prospects.Quadrupeds can jump any distance within their capabilities.Complex obstacles can be crossed by jumping.But the current quadruped robots are still far from having such flexible movement ability.To address this problem,a control framework based on deep reinforcement learning is proposed in this paper to realize the distance-controllable long jump of the robot.The control framework proposes a control strategy to achieve jumping by drawing on the jumping action of quadrupeds.Then deep reinforcement learning is used to learn the jumping parameters to realize the jumping control and distancecontrollable long jump control of the quadruped robot under different target distances.And the feasibility of the proposed method is verified by simulation and experiment.The specific work of the paper is as follows.(1)Kinematic and dynamical modeling of a quadruped robot.The mechanism configuration of the quadruped robot is analyzed,and its forward and inverse kinematic models are established.By calculating the Jacobi matrix,the mapping between the joint angular velocity and the linear velocity of the foot end is established.By calculating the force Jacobi,the mapping between the joint moment and the foot-end reaction force is established.Establish a simplified kinetic model of quadruped robot jumping,and design the robot jumping action for control frame design by combining the jumping process of quadrupeds.(2)Designing a framework for jump control of a quadruped robot.In this paper,we design a control framework for the jumping process of a quadruped robot,which includes a body trajectory planning module,a foot trajectory planning module,a touchdown detection module and a landing control module.The finite state machine is designed according to the change of the robot’s state during the jumping process,and can call the corresponding module in the framework to control the robot with the change of the robot’s state.(3)Deep reinforcement learning algorithms are introduced to optimize the jump control parameters.The coupling relationship between robot control parameters and body motion,parameters and parameters are complex and high-dimensional,and it is very difficult to solve the optimal control parameters.To address this kind of problem,this paper uses deep reinforcement learning algorithm to optimize the control parameters with a priori knowledge to realize the robot to complete the jumping task with different target distances under different starting states.(4)Simulation and experimental.A virtual robot of the quadruped robot is built in Gazebo to realize single-step jump control and continuous jump control of the quadruped robot at different target distances.The proposed control method is applied to the quadruped robot to verify the feasibility of the algorithm.